Dual Monitoring Communication for Self-Aware Network-on-Chip: Architecture and Case Study

Introduction

With the sizes of transistors constantly decreasing, modern chips are able to integrate a large number of processing cores. For instance, Vangal et al. (2008) presented an 80-core processor, and Truong et al. (2009) presented a 167-processor computing platform. In the near future, on-chip parallel processing will be running on 1000-core chips (Asanovic et al., 2006). Clearly, how to utilize this large amount of processing power becomes an important topic for integrated circuit and communication system design.

The dynamics of massively parallel on-chip computation platforms pose major challenges to the development of such systems. For one thing, parallel on-chip systems may need to run different applications, with various performance requirements. The details of applications are likely to be unknown at the design time. Secondly, a diversity of software might run on the system, which is likely to have different resource management schemes. Last but not least, the status of hardware becomes anything but predictable with the small transistor size. Variations (Miranda et al., 2009), for instance, significantly influence the hardware performance. Variations refer to the deviation of transistor sizes or metal wire dimensions from the nominal values. With considerable hardware variations, the system performance, such as speed and power consumption, and system dependability become dynamic and unpredictable.

Autonomic computing (Horn, 2001) is a promising paradigm to cope with the dynamics of massively parallel processing. Self-adaptive computing belongs to the broad spectrum of autonomic computing, but has a narrower conceptual scope. Paraphrased from Salehie and Tahvildari (2009), a self-adaptive system evaluates its own behavior and changes the behavior when the evaluation indicates failure to achieve functional goals or performance optimization is achievable.

This article addresses a specific aspect of self-adaptive computing: self-awareness. As the name itself suggests, a self-aware system is aware of its internal state (Hinchey & Sterritt, 2006). Self-awareness, as an essential part of self-adaptive computing, ensures the visibility of system status, based on which adaptation and optimization can be performed. To provide self-awareness, monitoring and diagnostic operations shall be designed in the system, which can dynamically trace the performance parameters and failures. In embedded system domain, common operations include the monitoring of execution speed, power and energy, errors and failures, as well as temperature.

The work proposes a DMC (dual monitoring communication) architecture to support self-awareness in Network-on-Chips (NoCs). NoC is a scalable communication architecture for massively parallel, high-performance on-chip processing (Jantsch & Tenhunen, 2003; Vangal et al., 2008). Components are modularized as nodes in a network, which supports communication between parallel processing units. DMC provides the communication infrastructure for diagnostic and monitoring operations on NoCs. One monitoring communication is integrated with data channel, to trace the run-time profile, e.g., data value and latency, of data communication. The other monitoring communication is separate from the data channel, and is used to report monitoring information and transmit reconfiguration commands if necessary. DMC architecture is not dependent on network topologies and sizes, flow control schemes or other adaptive techniques, thus it is scalable and generically applicable to different NoC architectures.